IMB-P2P Benchmarks

IMB-P2P Benchmarks#

The IMB-P2P component of the Intel(R) MPI Benchmarks provides shared memory transport oriented benchmarks for MPI-1 point-to-point communications.

The shared memory transport oriented benchmark has the following key features:

  • write payload feature: cache lines in the user-provided send-buffer must be touched before sending a message

  • read payload feature: cache lines in the user-provided recv-buffer must be touched after receiving a message

This is the only way to make adequate performance measurements of the shared memory transport implementation performance.

The payload, which is not recorded to the send-buffer before sending a message, may give an abnormal fake advantages to CMA-based memory copy function implementations.

The payload, which is not read from the recv-buffer after receiving a message, may give an abnormal fake advantages to implementations based on non-temporal message copying to the user-defined recv-buffer.

Although IMB-P2P is shared memory transport oriented benchmarks, it can also be run on several nodes.

The IMB-P2P component contains the following benchmarks:

See Also

Command-Line Control for IMB-P2P Benchmarks